BERT-BiLSMT-CRF-NER

Tensorflow solution of NER task Using BiLSTM-CRF model with Google BERT Fine-tuning

使用谷歌的BERT模型在BLSTM-CRF模型上进行预训练用于中文命名实体识别的Tensorflow代码'

中文文档请查看https://blog.csdn.net/macanv/article/details/85684284 如果对您有帮助，麻烦点个star,谢谢~~

Welcome to star this repository!

The Chinese training data($PATH/NERdata/) come from:https://github.com/zjy-ucas/ChineseNER

The CoNLL-2003 data($PATH/NERdata/ori/) come from:https://github.com/kyzhouhzau/BERT-NER

The evaluation codes come from:https://github.com/guillaumegenthial/tf_metrics/blob/master/tf_metrics/__init__.py

Try to implement NER work based on google's BERT code and BiLSTM-CRF network!

How to train

1. Download BERT chinese model :

wget https://storage.googleapis.com/bert_models/2018_11_03/chinese_L-12_H-768_A-12.zip

2. create output dir

create output path in project path:

mkdir output

3. Train model

first method

  python3 bert_lstm_ner.py   \
                  --task_name="NER"  \ 
                  --do_train=True   \
                  --do_eval=True   \
                  --do_predict=True
                  --data_dir=NERdata   \
                  --vocab_file=checkpoint/vocab.txt  \ 
                  --bert_config_file=checkpoint/bert_config.json \  
                  --init_checkpoint=checkpoint/bert_model.ckpt   \
                  --max_seq_length=128   \
                  --train_batch_size=32   \
                  --learning_rate=2e-5   \
                  --num_train_epochs=3.0   \
                  --output_dir=./output/result_dir/

OR replace the BERT path and project path in bert_lstm_ner.py

if os.name == 'nt': #windows path config
   bert_path = '{your BERT model path}'
   root_path = '{project path}'
else: # linux path config
   bert_path = '{your BERT model path}'
   root_path = '{project path}'

Than Run:

python3 bert_lstm_ner.py

USING BLSTM-CRF OR ONLY CRF FOR DECODE!

Just alter bert_lstm_ner.py line of 450, the params of the function of add_blstm_crf_layer: crf_only=True or False

ONLY CRF output layer:

    blstm_crf = BLSTM_CRF(embedded_chars=embedding, hidden_unit=FLAGS.lstm_size, cell_type=FLAGS.cell, num_layers=FLAGS.num_layers,
                          dropout_rate=FLAGS.droupout_rate, initializers=initializers, num_labels=num_labels,
                          seq_length=max_seq_length, labels=labels, lengths=lengths, is_training=is_training)
    rst = blstm_crf.add_blstm_crf_layer(crf_only=True)

BiLSTM with CRF output layer

    blstm_crf = BLSTM_CRF(embedded_chars=embedding, hidden_unit=FLAGS.lstm_size, cell_type=FLAGS.cell, num_layers=FLAGS.num_layers,
                          dropout_rate=FLAGS.droupout_rate, initializers=initializers, num_labels=num_labels,
                          seq_length=max_seq_length, labels=labels, lengths=lengths, is_training=is_training)
    rst = blstm_crf.add_blstm_crf_layer(crf_only=False)

Result:

all params using default

In dev data set:

In test data set

entity leval result:

last two result are label level result, the entitly level result in code of line 796-798,this result will be output in predict process. show my entity level result :

my model can download from baidu cloud:
链接：https://pan.baidu.com/s/1GfDFleCcTv5393ufBYdgqQ 提取码：4cus

ONLINE PREDICT

If model is train finished, just run

python3 terminal_predict.py

Using yourself data to train

if you want to use yourself data to train ner model,you just modify the get_labes func.

def get_labels(self):
       return ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC", "X", "[CLS]", "[SEP]"]

NOTE: "X", “[CLS]”, “[SEP]” These three are necessary, you just replace your data label to this return list.
Or you can use last code lets the program automatically get the label from training data

def get_labels(self):
        # 通过读取train文件获取标签的方法会出现一定的风险。
        if os.path.exists(os.path.join(FLAGS.output_dir, 'label_list.pkl')):
            with codecs.open(os.path.join(FLAGS.output_dir, 'label_list.pkl'), 'rb') as rf:
                self.labels = pickle.load(rf)
        else:
            if len(self.labels) > 0:
                self.labels = self.labels.union(set(["X", "[CLS]", "[SEP]"]))
                with codecs.open(os.path.join(FLAGS.output_dir, 'label_list.pkl'), 'wb') as rf:
                    pickle.dump(self.labels, rf)
            else:
                self.labels = ["O", 'B-TIM', 'I-TIM', "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC", "X", "[CLS]", "[SEP]"]
        return self.labels

NEW UPDATE

2019.1.9: Add code to remove the adam related parameters in the model, and reduce the size of the model file from 1.3GB to 400MB.

2019.1.3: Add online predict code

reference:

The evaluation codes come from:https://github.com/guillaumegenthial/tf_metrics/blob/master/tf_metrics/__init__.py
https://github.com/google-research/bert
https://github.com/kyzhouhzau/BERT-NER
https://github.com/zjy-ucas/ChineseNER

Any problem please email me(ma_cancan@163.com)

SusannaWull/BERT-BiLSTM-CRF-NER